CN111355936A

CN111355936A - Method and system for acquiring and processing image data for artificial intelligence

Info

Publication number: CN111355936A
Application number: CN201911315503.7A
Authority: CN
Inventors: 张光斌; 熊伟华; 熊智斌
Original assignee: Hangzhou Jingmou Intelligent Technology Co ltd
Current assignee: Hangzhou Jingmou Intelligent Technology Co ltd
Priority date: 2018-12-20
Filing date: 2019-12-19
Publication date: 2020-06-30
Anticipated expiration: 2039-12-19
Also published as: CN111355936B

Abstract

An intelligent vision processing device performs AI processing directly using RAW data from an image sensor, rather than after ISP processing. The details of the processing flow and the preprocessing method of the raw data are described and compared with the conventional method. Alternative parallel ISP paths may be used to generate data suitable for display. The invention enables better results to be obtained at lower cost and better performance.

Description

Method and system for acquiring and processing image data for artificial intelligence

Technical Field

The invention relates to a technology in the field of machine vision, in particular to a method and a system for acquiring and processing image data by artificial intelligence, which are suitable for equipment for processing visual information by the artificial intelligence.

Background

Vision is one of the most important ways to obtain information, both in the human world and in the machine world. In today's machine vision world, most systems use image sensors as front-ends. The RAW sensor output is typically in RAW format and does not match the color response of the human eye. In order to be pleasant to the human eye, an ISP (image signal processing) is used to convert raw data into color data, such as RGB data, suitable for monitor display and viewing by the human eye. Unfortunately, ISP switching may lose some useful information, introduce some erroneous information, and add more redundant data to the data set. As a result, the complexity of the subsequent AI processing units increases and their performance decreases.

Disclosure of Invention

In view of the above-mentioned shortcomings of the prior art, the present invention provides a method and system for acquiring and processing image data with artificial intelligence, which directly uses raw data from a sensor for preprocessing, and obtains better results with lower cost and better performance.

The invention is realized by the following technical scheme:

the invention relates to a method for collecting and processing image data for artificial intelligence, which separates the original data collected by a semiconductor device into a plurality of single-channel images according to channels, judges the information content of the separated single-channel images, and performs signal enhancement and AI processing on the images with the most information content.

The information amount decision can contain the optimal original information depending on the image of a single channel or the combined image of a plurality of channels.

The invention relates to a system for acquiring and processing image data for artificial intelligence, comprising: the system comprises an image sensor used for collecting original visual data, a preprocessing unit connected with the image sensor and used for optimizing the original visual data, an AI unit used for processing the visual data, and a parallel ISP adjustable path arranged between the preprocessing unit and the AI unit.

The preprocessing unit adjusts the original visual data obtained by the image sensor into preprocessing information which is more suitable for AI processing and maintains the format of the original data.

The system is preferably realized in the mode of an integrated circuit chip or a chip set; further preferably one or more silicon chips for efficient edge computing applications.

Drawings

FIG. 1 is a block diagram of an AI visual process;

FIG. 2 is a flow chart of the present invention;

FIG. 3 is a flow chart of the ISP converting sensor raw data to RGB;

FIG. 4 is a schematic diagram of a pretreatment unit;

FIG. 5 is a schematic diagram of a channel selection combiner;

fig. 6 is a schematic diagram of a signal enhancer and a bit compressor.

Detailed Description

As shown in fig. 1, for AI (artificial intelligence) flow for vision applications, RAW data is collected by an image sensor 100 in RAW format, processed by an ISP (image signal processing) unit 110 to generate color images, typically in RGB format (R: red, G: green, B: blue), and then passed to a display unit 120, such as an LED display, for display. The main purpose of the ISP and display is to make the eye pleasing. In recent years, the acquired image data is no longer used only for human eyes but also for analysis of AI, thereby motivating various machine vision-based applications such as automatic driving, intelligent robots, intelligent monitoring, and the like. The output of the ISP is also used as an input to AI unit 130 in most current cases. The functions of AI element 130 may include, but are not limited to, object detection, searching, indexing, identification, and the like. Unfortunately, the flow of the AI process is not optimized because the main purpose of the ISP is to better show it.

As shown in fig. 2, the process flow related to the present embodiment for optimizing AI further includes, on the basis of the selectable path, that is, the raw data is collected by the image sensor 200 and displayed by the ISP 210 and the display module 220: the raw data is collected by the image sensor 220 and output to the preprocessing unit 230 and output to the display unit 220 and the subsequent stage of the machine vision-based application program after passing through the AI processing unit 240, respectively, as some additional or enhanced information for enhanced display, wherein: the input and output of the preprocessing unit 230, i.e. the raw data raw1 and the preprocessed data raw2, are in raw format.

As shown in fig. 3, is a process of passing raw imaging data to an AI. The raw image data is typically in a raw format 300 in a Bayer Pattern (Bayer Pattern). It has only one color at each pixel location. There is one red pixel, one blue pixel and two green pixels in each repeating Bayer image cell. After the ISP, the data format 310 is typically converted to RGB format, which has three color channel component values of red, green and blue at each pixel location simultaneously. In some cases, data format 310 may also be other ISP processed data formats such as YUV, but these ISP processed data formats have in common that each pixel contains all color channel information. For this purpose, a CIP (color interpolation process module) should be included in the ISP. The ISP may also include, but is not limited to, other color processing such as gamma adjustment, contrast adjustment, edge enhancement, noise reduction, tone mapping, auto black level, etc. The output RGB data is then transmitted to the AI unit 320. The algorithm is typically optimized for human vision and is not preferred for the AI process.

As shown in fig. 4, the raw data raw1 may be a bayer format image 400 or other color image of raw data acquired by a typical CMOS image sensor, and is separated into single-

channel images

420, 430, 440 by a channel separator 410, i.e., a sub-raw1 for the G channel, a sub-raw2 for the B channel, and a sub-raw3 for the R channel.

Notably, in the Bell format mode, the data count of the G channel is twice that of each of the other channels.

The separated

single channel images

420, 430, 440 are sent to a channel selection combiner 450 to select the most informative data set to pass to a signal enhancer 460 for better signal for subsequent AI processing, preferably with a bit compressor 470 after enhancement to further reduce data bandwidth.

The pixel format output by the signal enhancer 460 can be, but is not limited to, set to 8 bits; the bit compressor 470 for that reduces the pixel format from 8 bits to 4 bits without losing any critical information.

As shown in fig. 5, the channel selection combiner 450 receives a single-channel image, i.e., the separated

original sub-channel data

500, 510, and 520, which correspondingly contain information amounts satisfying the functions f1(G), f2(B), and f3 (R).

The function uses the subchannel RAW data as an input array and outputs a uniform value of how much information it contains.

The functions may be the same, i.e., f1 ═ f2 ═ f3, except for the input values; or the functions are different, i.e. f1, f2, f3 are not equal, e.g. entropy functions

Wherein p (x) is the probability of occurrence of x; or a derivative function thereof may be used as the fn () function.

The channel selection combiner 450 judges the values of the three functions according to the judgment units 530 to 590, and selects a corresponding single-channel image or a combination of a plurality of channel images according to the judgment result to output to the AI unit 595, for example, whether ① f1> (f2+ f3)/2, ② f2> f3/2 and ③ f3> f2/2 are met in sequence, and correspondingly takes the image of G, B, R or B + R channel as the input of the AI unit 595.

As shown in fig. 6, the present embodiment takes a popular "Lenna" image as an example, when 5 ROIs (regions of interest) are included in the single channel image 600, since the exposure and gain of the image sensor are usually optimized for the whole viewing area, it is not always well optimized for the local area. As shown in the sub-image group 610, one ideal situation and four non-ideal situations are included, where the non-ideal image quality includes too bright, too dark, too noisy, or low contrast, etc. By the local signal enhancement processing, the sub-image group can be digitally adjusted to be close to an ideal state.

When the signal enhancer 460 outputs the enhanced 8-bit image block 620, it is preferable to cut the LSB and reduce the number of bits to 4 bits by the bit compressor 470 and output the 4-bit image block 630 to the AI unit 640 for subsequent processing.

In the above embodiment, 8 bits of each color channel image data group is one typical value. In practical cases, the number of bits may vary depending on the raw data from the image sensor, which may be a lower or higher number of bits, e.g. 1bit, 2bit, 8bit, 10bit, 12bit, 14bit, 16bit, etc., depending on the dynamic range and data format of the sensor.

The above description uses a Bayer pattern to RGB conversion as an ISP example. In practice, the raw data may be in a format other than a Bayer image, such as an RGBC image, an RGB/IR image, or the like. The ISP output format may be other color images such as YUV, CMY, etc.

The invention can obtain the effects by the method, which comprises the following steps:

1. lower cost, lower power consumption, higher speed: comparing the technique shown in fig. 3 with the inventive method shown in fig. 4 for the same size of image input, the signal bandwidth can be reduced to 12 to 24 times lower than the conventional AI process flow after the ISP. This is highly desirable in hardware implementations, such as NPU, GPU, FPGA, ASIC silicon chips. The complexity of the corresponding deep learning network may also be reduced, since the overall data volume and size of the process is greatly reduced compared to conventional RGB data sets. As a result, lower chip cost and lower power consumption can be achieved. Or higher system processing speeds can be achieved with the same cost and power consumption budget.

2. The method also contributes to achieving better performance of AI processing: first, conventional ISP flows reduce useful information. For example. Typically, RAW data is 10 or 12 bits, even 14 bits for standard sensors and even higher for HDR sensors. In a conventional ISP, data is typically reduced to only 8 bits, since the purpose of monitor display and printing is to make the eye look nice. As a result, AI processing after the ISP can only obtain 8-bit data as input. This is not very good for many high contrast scenes, but is even more so for some high dynamic range scenes. Useful information from viewing objects located in relatively dark or bright areas of the image is often lost. As another example, in order for an image to look better in low light conditions, conventional ISPs typically use a powerful noise filtering function. Thereafter, the image may look much more beautiful, but a large amount of useful detailed information is lost. The subsequent AI cannot recover any detail information that has been lost in the original picture anyway. The invention is based on the AI processing flow of the original data, directly uses the data before the ISP as the input, can enable the data bit number processed by the AI unit to reach 10 bits or more, and simultaneously, the noise filtering function can pertinently preprocess the image, which is helpful for extracting useful information for the subsequent AI processing unit.

Second, the conventional ISP procedure may introduce some false information. As an example, the process flow of changing from a bayer format image to a color image typically involves a color interpolation process, a so-called "demosaicing" process. Where the three-color representation is reconstructed by estimating the missing components of neighboring pixels. In many cases, the estimated data is incorrect, color aliasing can result due to crosstalk of neighboring pixel data, and more color noise can also be introduced. As another example, edge enhancement is commonly used in ISPs to make images appear sharper. It will appear clearer but will actually produce an abrupt signal strength change at the edge of the target object image that is not present, which may lead to a false understanding by the AI processor. In contrast, the method uses the raw data directly without guessing the points from the ISP demosaicing algorithm and without intentionally enhancing the edges of the display. As a result, the AI process is based on more reliable data with higher confidence.

In addition, an original signal preprocessing unit can be further added in the flow according to needs to optimize data, so that the data is more suitable for AI processing and the AI performance is further improved.

The foregoing embodiments may be modified in many different ways by those skilled in the art without departing from the spirit and scope of the invention, which is defined by the appended claims and all changes that come within the meaning and range of equivalency of the claims are therefore intended to be embraced therein.

Claims

1. A method for collecting and processing image data used for artificial intelligence is characterized in that original data collected by a semiconductor device are separated into a plurality of single-channel images according to channels, the separated single-channel images are subjected to information quantity judgment, and the image with the largest information quantity is subjected to signal enhancement and AI processing.

2. The method of claim 1, wherein said information content decision is based on whether a single channel image or a multi-channel combined image can contain optimal original information.

3. The method for artificial intelligence acquisition and processing of image data as recited in claim 1, wherein a bit compressor is employed after enhancement to further reduce data bandwidth.

4. A system for acquiring and processing image data for artificial intelligence, comprising: the system comprises an image sensor for acquiring original visual data, a preprocessing unit connected with the image sensor and used for optimizing the original visual data, an AI unit for processing the visual data, and a parallel ISP adjustable path arranged between the preprocessing unit and the AI unit, wherein: the input and output of the preprocessing unit, i.e. both raw data and preprocessed data, are in raw format.

5. The system of claim 4, wherein the image sensor collects raw data and outputs the raw data to the preprocessing unit and the AI processing unit and then outputs the raw data to the display unit and the subsequent stage of the machine vision based application program, respectively, as some additional or enhanced information for enhanced display.

6. The system of claim 4, wherein the channel selection combiner receives a single-channel image, i.e. the separated original sub-channel data, performs the judgment according to the information amounts f1(G), f2(B) and f3(R) contained therein, and selects the corresponding single-channel image or the combination of multiple channel images to output to the AI unit according to the judgment result.

7. The system of claim 6, wherein the channel selection combiner sequentially determines whether ① f1> (f2+ f3)/2, ② f2> f3/2, ③ f3> f2/2 are satisfied, and corresponds to the image of G, B, R or B + R channel as the input of AI unit.

8. The system of claim 4, wherein the system is implemented in the form of an integrated circuit chip or chipset, in particular one or more silicon chips for efficient edge computing applications.